Releases: facebook/prophet
Releases · facebook/prophet
1.1.5
What's Changed
Python
- Upgraded cmdstan version to 2.33.1, enabling Apple M2 support. @lucentcosmos @tcuongd
- Added a binary distribution for the macOS arm64 architecture (M1, M2 chips). @tcuongd
- Added argument
scaling
to theProphet()
instantiation. Allowsminmax
scaling ony
instead of
absmax
scaling (dividing by the maximum value).scaling='absmax'
by default, preserving the
behaviour of previous versions. @yoziru - Added argument
holidays_mode
to theProphet()
instantiation. Allows holidays regressors to have
a different mode than seasonality regressors.holidays_mode
takes the same value asseasonality_mode
if not specified, preserving the behaviour of previous versions. @CoreyBryant-everi - Added two methods to the
Prophet
object:preprocess()
andcalculate_initial_params()
. These
do not need to be called and will not change the model fitting process. Their purpose is to provide
clarity on the pre-processing steps taken (y
scaling, creating fourier series, regressor scaling,
setting changepoints, etc.) before the data is passed to the stan model. @tcuongd - Added argument
extra_output_columns
tocross_validation()
. The user can specify additional columns
frompredict()
to include in the final output alongsideds
andyhat
, for exampleextra_output_columns=['trend']
. @dchiang00 - prophet's custom
hdays
module was deprecated last version and is now removed.
R
- Updated holidays data based on holidays v0.34
Full Changelog: v1.1.4...1.1.5
1.1.4
What's Changed
Python
- We now rely solely on the
holidays
package for country holidays. Credits to @arkid15r in #2379- This allows us to take full advantage of improvements to the
holidays
package, and removes reliance on unmaintained manual holidays entries inhdays.py
. Importing from theprophet.hdays
module has been deprecated and the module will be removed in the next release. - holidays v0.20 and beyond contains most / all previously missing holidays, so there should be very little differences in fitted models and predictions that currently make use of inbuilt country holidays.
- Note that for countries like India and Pakistan, some Christian holidays are currently not included in the
holidays
package so will not be added automatically with.add_country_holidays()
. These can be added manually instead, see examples here.
- This allows us to take full advantage of improvements to the
- Upgraded underlying cmdstan to v2.31.0, which fixes installation issues on Apple M1. Credits to @WardBrian in #2428
- Fixed a bug with Windows wheel builds caused by long path names. Credits to @WardBrian
R
- Updated holidays data based on holidays v0.25.
v1.1.3-patched
What's Changed
Full Changelog: v1.1.2...v1.1.3-patched
1.1.2
What's Changed
Python
- Sped up
.predict()
by up to 10x by removing intermediate DataFrame creations. @orenmatar (#2299) - Sped up fourier series generation, leading to at least 1.5x speed improvement for
train()
andpredict()
pipelines. @yoziru (#2334) - Fixed bug in how warm start values were being read. Documentation has been updated. @tcuongd
- Developer experience: modernized build / develop / test workflow based on https://discuss.python.org/t/custom-build-steps-moving-bokeh-off-setup-py/16128/17 @tcuongd
- Wheels are now version-agnostic. This is possible because we don't use any Python-version-specific tooling to compile stan models. Credits to @WardBrian for the suggestion and this example: https://github.com/joerick/python-ctypes-package-sample.
R
- Fixed a bug in
construct_holiday_dataframe()
- Updated
holidays
data based on holidays version 0.18. - Note that the 1.1.2 has been submitted to CRAN but is not live yet. You can install by downloading the attached files:
.tar.gz
to install from source, or.tgz
for the macOS binary.
1.1.1
What's Changed
Python
- Improved runtime of
predict()
function via vectorization of future draws. Details here. Credits to @orenmatar for the original blog post and @winedarksea @tcuongd for the implementation.predict()
now has a new argument,vectorized
, which is true by default. You should see speedups of 3-7x for predictions, especially if the model does not use full MCMC sampling. When usinggrowth='logistic'
withmcmc_samples > 0
, predictions may be slower, and in these cases you can fall back to the original code by specifyingvectorized=False
.
- Added aarch64 wheels for Linux and parallelised wheel build workflow. @thechopkins
cmdstanpy
minimum version is now 1.0.4.- Fixed a bug where the version number hadn't updated from 1.0.
prophet.__version__
now returns the correct version. @tcuongd
R
- (Backend change) Make holidays data internal to the package to prevent unintentional overrides. Note that the data can still be read by end users as before, this hasn't changed. @bartekch
1.1
What's Changed
Python
- Minimum required version of Python is now 3.7
- Removed dependency on
pystan==2.19.1.1
, which is no longer maintained.cmdstanpy
is now the sole stan backend. @tcuongd @WardBrian @akosfurton @malmashhadani-88 - Python binaries built for MacOS, Linux, and Windows for Python 3.7-3.10, so end users no longer need to compile the Prophet model from source on their machine. @tcuongd
- Binaries are built using Github Actions. @tcuongd @abitrolly
- Use normal-id-glm distribution for Stan model to improve MCMC sampling speed (the model itself is still identical). @andrjohns
- Improved execution time of
rolling_mean_by_h
function used to calculate cross validation performance metrics. @RaymondMcT
R
- Update holidays data based on
holidays
package version 0.13.
v1.0
v0.7
v0.6
- Fix bugs related to upstream changes in
holidays
andpandas
packages. - Compile model during first use, not during install (to comply with CRAN policy)
cmdstanpy
backend now available in Python